Recent publications on ArXiv dated July 2026 shed light on both the progress and persistent challenges in the reliability, efficiency, and privacy of artificial intelligence systems, particularly generative ones. These studies underscore the importance of a rigorous approach to ensure that AI is not only powerful but also controllable and accountable, a fundamental principle for ethical AI.
What happened
Several research teams have published significant works. One study revealed that LLM-as-a-Judge scores, meaning the use of Large Language Models to self-evaluate their own performance, proved to be unreliable optimization signals in closed-loop table recognition. The research, published in LLM-as-a-Judge Scores Are Unreliable Optimization Signals in Closed-Loop Table Recognition, showed that scores frequently tied, rankings were not reproducible, and no tested evaluation policy significantly improved performance, calling into question a widely adopted feedback method.
In another direction, Generative Compilation was introduced as an innovative approach to improve AI code generation. As described in Generative Compilation: On-the-Fly Compiler Feedback as AI Generates Code, this method provides real-time feedback from a compiler during an LLM's code generation process. This is particularly useful for languages with strict static semantics like Rust, where compliance with language rules is critical. Immediate feedback guides the AI to produce more correct and robust code, overcoming the limitations of methods that intervene only after generation is complete.
Continuing on the theme of software reliability, another study presented Monty, a framework for faithful autoformalization of natural language assertions. The work Faithful Autoformalization of Natural Language Assertions aims to bridge the gap between informal developer intent and formal executable specifications, essential for software testing and verification. Monty addresses challenges related to assertion validity and natural language ambiguity, making the process of creating formal contracts more efficient and less error-prone.
Finally, to address the practical challenges of LLM deployment, an efficient and privacy-aware edge-cloud collaborative inference framework was proposed. This system, detailed in Efficient and Privacy Aware Edge Cloud Collaborative Inference for Large Language Models, distributes the computational load between local edge devices and the cloud. The goal is to balance response latency, limited hardware resources on devices, and user privacy, ensuring that sensitive data is not exposed to the cloud—a crucial aspect for the widespread adoption of LLMs in consumer and enterprise contexts.
Why it matters
These developments have profound implications for how we design, validate, and deploy AI. The discovery of LLMs' unreliability as self-judges raises serious questions about the validity of many closed-loop optimization systems and highlights the need for external, objective evaluation metrics. Trust in generative AI systems depends on their ability to be independently measured and verified.
On the other hand, innovations in Generative Compilation and autoformalization demonstrate how AI can be made more effective and less prone to errors, especially in complex tasks like code generation and verification. This can accelerate software development and improve quality, but it also requires developers to understand the limitations and responsibilities associated with using AI tools. The balance between automation and human oversight becomes increasingly critical.
The edge-cloud collaborative inference framework is fundamental for the large-scale adoption of LLMs. It addresses the "trilemma" of performance, resources, and privacy, making AI more accessible without compromising personal data protection. This is vital for public and regulatory acceptance of AI, especially in Europe with regulations like GDPR and the upcoming EU AI Act.
The HDAI perspective
The presented research reinforces the vision of Human Driven AI: artificial intelligence must be developed and deployed with a primary focus on human beings, their needs, and their safety. The unreliability of LLMs as self-judges reminds us that human validation and independent auditing remain irreplaceable. It is not just a matter of technical efficiency, but of trust and human control. Advancements in code generation and autoformalization, while promising, require robust AI governance to prevent errors and ensure accountability.
The issue of privacy in edge-cloud inference is a clear example of how architectural choices directly impact fundamental rights. It is essential that technological innovation is always accompanied by deep ethical reflection and a clear regulatory framework. These topics will be central to discussions at the HDAI Summit 2026, where experts from around the world will debate how to build an AI future that is ethical, responsible, and serves humanity.
What to watch
It will be crucial to observe how the research community responds to the evidence on the unreliability of LLM-as-a-Judge, developing more robust evaluation metrics and external verification systems. In parallel, the evolution of techniques like Generative Compilation and autoformalization will continue to redefine the role of developers, who will need to acquire new skills to collaborate effectively with AI. Finally, the practical implementation of distributed and privacy-aware inference solutions will be a key indicator of AI's maturity in enterprise and consumer contexts, with an eye on Italian AI innovation and its ability to integrate these principles.